AgentVQA Games
Multimodal · 2026-10-22
The 200 AgentVQA questions in the Games domain, covering game-state, scene-understanding and next-action questions from video games, Atari and Super Mario gameplay. Questions are manually verified four-option multiple-choice items converted from 23 existing datasets, so offline answers stand in for online agent actions such as choosing the next GUI click or predicting a robot's next step. Models answer with a single letter from the visual input (images, screenshots or video) and any action history; chance is 25%. The score is the percentage of these questions answered correctly.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.7 Flash | 70.0 |
| Gemini 3.1 Pro Preview | 68.0 |
| Kimi K3 | 63.5 |
| GPT-5.6 Sol | 62.5 |
| Qwen3.7-Plus (source unspecified) | 61.5 |
| Qwen3.8 27B | 61.5 |
| GPT-5.6 Luna | 61.0 |
| Qwen3.7 Flash | 61.0 |
| Qwen3.5 397B A17B | 57.5 |
| Qwen3.5 35B A3B | 55.5 |
| Sonnet 5 | 53.0 |
| Qwen3.5 122B A10B | 53.0 |
| Opus 4.8 | 50.5 |
| Qwen3.5-27B | 49.5 |
| GLM-4.6V | 47.5 |
| GLM-4.5V | 46.5 |
| GLM-4.1V-9B-Thinking | 41.0 |
| GPT-4o (source unspecified) | 29.0 |
| Llama 4 Maverick Instruct | 23.0 |